./run_clm_flax.py \
    --output_dir="./" \
    --model_type="gpt2" \
    --model_name_or_path="./" \
    --config_name="./" \
    --tokenizer_name="./" \
    --dataset_name="oscar" \
    --dataset_config_name="unshuffled_deduplicated_es" \
    --do_train \
    --do_eval \
    --block_size="512" \
    --per_device_train_batch_size="64" \
    --per_device_eval_batch_size="64" \
    --adafactor \
    --learning_rate="3e-3" --warmup_steps="1000" \
    --adam_beta1="0.9" --adam_beta2="0.98" --weight_decay="0.001" \
    --overwrite_output_dir \
    --num_train_epochs="20" \
    --logging_steps="100" \
    --save_steps="2500" \
    --eval_steps="1000000" \
    --preprocessing_num_workers="64"